摘要
Deepfake videos threaten digital media integrity, and existing detectors struggle to jointly capture both spatial artifacts (e.g., unnatural skin textures) and temporal artifacts (e.g., inconsistent facial dynamics). They also average clip predictions, ignoring dramatic segment-level forgery saliency variation. Therefore, we introduce STEN-FAWA, a Spatial-Temporal Expert Network with a Forgery-Aware Weighted-Adaptive Aggregator. STEN first uses separate vision and video Transformers, specializing in spatial and temporal artifacts, respectively. FAWA then generates adaptive weights for each clip's prediction to amplify the suspicious one. To mimic real-world forgeries, we further propose PFSA, which synthesizes sparsely manipulated videos with random-intensity and random-region artifacts. Under the most challenging unseen-dataset protocol, STEN-FAWA achieves an AUC of85.18 %on DFDC, surpassing prior state-of-the-art methods by up to3.38 % . These results underscore the benefits of expert specialization, adaptive aggregation, and realistic augmentation for robust deepfake detection.